Ranking run 2026-07-11
Claude Opus 4.7 leads with 79/100 across all domains. Tax law: Claude Opus 4.7 (75) · Medicine: Claude Opus 4.7 (88) · Law: Claude Opus 4.7 (80) · Business law: GPT-5 (81).
Overall across domains
| # | Model | top100 | 95 % CI | Δ prev. | failures | sec median/p90 | hallu band | per 1000 tok. | Δ own family |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.7 | 0 | 20 / 31 | 79 | 76–82 | -1 | 60 | 1,41 | — |
| 2 | GPT-5.6 Sol | 0 | 79 / 191 | 77 | 73–80 | — | 67 | 0,28 | — |
| 3 | Claude Opus 4.8 | 0 | 19 / 30 | 75 | 72–78 | 0 | 72 | 0,95 | — |
| 3 | GPT-5 | 0 | 31 / 46 | 75 | 71–79 | -1 | 70 | 0,60 | — |
| 4 | Kimi K2.6 n=78/100 | 22 (22 T) | 169 / 249 | 58 | 53–63 | -1 | 48 | 0,27 | — |
| 5 | Grok 4.5 | 0 | 16 / 31 | 56 | 52–61 | — | 72 | 3,99 | — |
| 6 | Gemini 3.1 Pro | 0 | 20 / 28 | 55 | 51–59 | -1 | 74 | 4,34 | — |
| 7 | Mistral Large 2 | 0 | 9 / 16 | 51 | 47–56 | 0 | 53 | 4,15 | — |
| 8 | GLM-4.6 n=97/100 | 3 (3 T) | 61 / 86 | 45 | 40–49 | -2 | 65 | 0,48 | — |
Δ prev. = difference to the previous published run; “—” = model was not in it. Hallu band: 100 = none, 0 = ≥ 4 false claims per answer. Per 1000 tok. = flagged claims per 1000 output tokens (length-normalised). Δ own family = score with all judges minus score with foreign judges only, in points; positive = own family scores more generously; “—” = not recorded before code_version 0.2.8 or no foreign judge (methodology §4, §6.5).
Tax law · 15 Items
Ranks 1–4 statistically tied (95 % CI overlaps rank 1)
In Tax law, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,4 times per answer and adds avg 1,5 claims the judge panel flags as false or unsupported; 100 % of answers carry at least one flag. The lowest value is Gemini 3.1 Pro at avg 1,1 per answer. Relative to answer length: 1,54 per 1000 output tokens, lowest GPT-5.6 Sol at 0,29.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.7 | 75 | 67–82 | +2 | 0,4 | 1,5 | 1,9 | 100 % | 1,54 | 75 % |
| 2 | GPT-5.6 Sol | 70 | 61–78 | — | 1,1 | 1,4 | 2,5 | 100 % | 0,29 | 69 % |
| 3 | Claude Opus 4.8 | 67 | 62–73 | +1 | 0,7 | 0,8 | 1,6 | 100 % | 1,12 | 68 % |
| 4 | GPT-5 | 65 | 56–74 | +5 | 1,0 | 0,8 | 1,8 | 87 % | 0,65 | 73 % |
| 5 | Gemini 3.1 Pro | 48 | 41–55 | 0 | 0,9 | 0,3 | 1,1 | 73 % | 4,68 | 63 % |
| 6 | Grok 4.5 | 47 | 38–55 | — | 1,1 | 0,7 | 1,7 | 93 % | 5,14 | 56 % |
| 7 | Kimi K2.6 | 33 | 19–45 | +2 | 1,9 | 1,5 | 3,4 | 100 % | 0,34 | 59 % |
| 8 | Mistral Large 2 | 28 | 21–35 | +1 | 1,7 | 1,1 | 2,8 | 100 % | 6,13 | 52 % |
| 9 | GLM-4.6 | 21 | 13–29 | 0 | 2,1 | 0,8 | 2,8 | 100 % | 0,89 | 64 % |
Medicine · 20 Items
Ranks 1–4 statistically tied (95 % CI overlaps rank 1)
In Medicine, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,3 times per answer and adds avg 0,8 claims the judge panel flags as false or unsupported; 90 % of answers carry at least one flag. The lowest value is Grok 4.5 at avg 0,3 per answer. Relative to answer length: 1,04 per 1000 output tokens, lowest GLM-4.6 at 0,13.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.7 | 88 | 83–92 | +4 | 0,3 | 0,8 | 1,1 | 90 % | 1,04 | 87 % |
| 2 | Claude Opus 4.8 | 85 | 81–89 | +5 | 0,4 | 0,4 | 0,6 | 85 % | 0,59 | 80 % |
| 3 | GPT-5.6 Sol | 84 | 79–88 | — | 0,5 | 0,3 | 0,7 | 80 % | 0,42 | 77 % |
| 4 | GPT-5 | 81 | 73–88 | -1 | 0,4 | 0,6 | 1,0 | 95 % | 0,70 | 79 % |
| 5 | Kimi K2.6 | 69 | 57–79 | -5 | 0,7 | 0,4 | 1,0 | 80 % | 0,21 | 74 % |
| 6 | Grok 4.5 | 67 | 58–76 | — | 0,3 | 0,1 | 0,3 | 55 % | 1,24 | 66 % |
| 7 | Mistral Large 2 | 67 | 60–73 | +2 | 0,5 | 0,6 | 1,0 | 100 % | 2,07 | 64 % |
| 8 | Gemini 3.1 Pro | 64 | 58–70 | -1 | 0,2 | 0,3 | 0,4 | 60 % | 1,67 | 70 % |
| 9 | GLM-4.6 | 57 | 52–64 | -2 | 0,1 | 0,2 | 0,3 | 40 % | 0,13 | 74 % |
Law · 35 Items
Ranks 1–4 statistically tied (95 % CI overlaps rank 1)
In Law, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,2 times per answer and adds avg 0,8 claims the judge panel flags as false or unsupported; 77 % of answers carry at least one flag. The lowest value is Claude Opus 4.8 at avg 0,6 per answer. Relative to answer length: 1,11 per 1000 output tokens, lowest Kimi K2.6 at 0,20.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.7 | 80 | 75–85 | -2 | 0,2 | 0,8 | 1,0 | 77 % | 1,11 | 79 % |
| 2 | Claude Opus 4.8 | 79 | 75–82 | -3 | 0,2 | 0,4 | 0,6 | 66 % | 0,70 | 78 % |
| 3 | GPT-5.6 Sol | 75 | 68–82 | — | 0,5 | 0,6 | 1,1 | 89 % | 0,31 | 77 % |
| 4 | GPT-5 | 71 | 63–78 | -3 | 0,6 | 0,6 | 1,1 | 86 % | 0,64 | 69 % |
| 5 | Gemini 3.1 Pro | 60 | 52–67 | -1 | 0,5 | 0,2 | 0,7 | 54 % | 4,03 | 70 % |
| 6 | Kimi K2.6 | 58 | 51–65 | -7 | 0,8 | 0,8 | 1,6 | 83 % | 0,20 | 64 % |
| 7 | Grok 4.5 | 52 | 44–60 | — | 0,9 | 0,3 | 1,1 | 63 % | 6,67 | 72 % |
| 8 | Mistral Large 2 | 49 | 40–57 | 0 | 0,9 | 0,8 | 1,7 | 77 % | 5,90 | 68 % |
| 9 | GLM-4.6 | 46 | 39–53 | -5 | 0,7 | 0,5 | 1,1 | 68 % | 0,42 | 59 % |
Business law · 30 Items
Ranks 1–4 statistically tied (95 % CI overlaps rank 1)
In Business law, GPT-5 (top score) contradicts a stored key fact on average 0,4 times per answer and adds avg 0,8 claims the judge panel flags as false or unsupported; 97 % of answers carry at least one flag. Relative to answer length: 0,50 per 1000 output tokens, lowest GPT-5.6 Sol at 0,22.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT-5 | 81 | 74–86 | -1 | 0,4 | 0,8 | 1,2 | 97 % | 0,50 | 74 % |
| 2 | GPT-5.6 Sol | 77 | 71–83 | — | 0,5 | 0,9 | 1,4 | 100 % | 0,22 | 68 % |
| 3 | Claude Opus 4.7 | 73 | 68–79 | -4 | 0,6 | 1,8 | 2,5 | 100 % | 1,74 | 66 % |
| 4 | Claude Opus 4.8 | 69 | 63–75 | 0 | 0,6 | 1,2 | 1,8 | 97 % | 1,21 | 70 % |
| 5 | Kimi K2.6 | 61 | 55–68 | +4 | 0,8 | 2,0 | 2,8 | 100 % | 0,35 | 55 % |
| 6 | Grok 4.5 | 59 | 50–67 | — | 0,7 | 0,7 | 1,3 | 80 % | 3,44 | 59 % |
| 7 | Mistral Large 2 | 56 | 49–62 | 0 | 0,8 | 1,3 | 2,1 | 97 % | 3,50 | 56 % |
| 8 | Gemini 3.1 Pro | 47 | 39–56 | -2 | 1,2 | 0,7 | 1,8 | 97 % | 5,67 | 52 % |
| 9 | GLM-4.6 | 47 | 40–54 | -1 | 0,8 | 1,0 | 1,7 | 100 % | 0,51 | 57 % |
Raw data of this run
raw.jsonl and progress.jsonl are the public streams with response texts stripped (contamination protection, methodology §7).